Journal of Machine Learning Research 18 (2017) 1-37 Submitted 11/16; Revised 6/17; Published 10/17 Stability of Controllers for Gaussian Process Dynamics Julia Vinogradska1,2
[email protected] Bastian Bischoff1
[email protected] Duy Nguyen-Tuong1
[email protected] Jan Peters2,3
[email protected] 1Corporate Research, Robert Bosch GmbH Robert-Bosch-Campus 1 71272 Renningen 2Intelligent Autonomous Systems Lab, Technische Universit¨atDarmstadt Hochschulstraße 10 64289 Darmstadt 3Max Planck Institute for Intelligent Systems Spemannstraße 10 72076 T¨ubingen Editor: George Konidaris Abstract Learning control has become an appealing alternative to the derivation of control laws based on classic control theory. However, a major shortcoming of learning control is the lack of performance guarantees which prevents its application in many real-world scenarios. As a step towards widespread deployment of learning control, we provide stability analysis tools for controllers acting on dynamics represented by Gaussian processes (GPs). We consider differentiable Markovian control policies and system dynamics given as (i) the mean of a GP, and (ii) the full GP distribution. For both cases, we analyze finite and infinite time horizons. Furthermore, we study the effect of disturbances on the stability results. Empirical evaluations on simulated benchmark problems support our theoretical results. Keywords: Stability, Reinforcement Learning, Control, Gaussian Process 1. Introduction Learning control based on Gaussian process (GP) forward models has become a viable approach in the machine learning and control theory communities. Many successful applica- tions impressively demonstrate the efficiency of this approach (Deisenroth et al., 2015; Pan and Theodorou, 2014; Klenske et al., 2013; Maciejowski and Yang, 2013; Nguyen-Tuong and Peters, 2011; Engel et al., 2006; Kocijan et al., 2004).